Exploring Probabilistic Soft Logic as a framework for integrating top-down and bottom-up processing of language in a task context

TL;DR

This paper presents a novel architecture that combines top-down semantic context with bottom-up linguistic analysis using Probabilistic Soft Logic to improve understanding and normalization of non-standard language input, demonstrated on German learner answers.

Contribution

The architecture integrates multiple NLP components into a probabilistic graphical model for joint analysis, enabling context-aware interpretation of non-standard language forms.

Findings

Effective integration of semantic context and linguistic analysis.

Framework adaptable to various languages and tasks.

Potential for improved normalization and understanding of learner language.

Abstract

This technical report describes a new prototype architecture designed to integrate top-down and bottom-up analysis of non-standard linguistic input, where a semantic model of the context of an utterance is used to guide the analysis of the non-standard surface forms, including their automated normalization in context. While the architecture is generally applicable, as a concrete use case of the architecture we target the generation of semantically-informed target hypotheses for answers written by German learners in response to reading comprehension questions, where the reading context and possible target answers are given. The architecture integrates existing NLP components to produce candidate analyses on eight levels of linguistic modeling, all of which are broken down into atomic statements and connected into a large graphical model using Probabilistic Soft Logic (PSL) as a framework. Maximum a posteriori inference on the resulting graphical model then assigns a belief distribution to candidate target hypotheses. The current version of the architecture builds on Universal Dependencies (UD) as its representation formalism on the form level and on Abstract Meaning Representations (AMRs) to represent semantic analyses of learner answers and the context information provided by the target answers. These general choices will make it comparatively straightforward to apply the architecture to other tasks and other languages.